Incorporating correlations between drugs and heterogeneity of multi-omics data in structured penalized regression for drug sensitivity prediction
Targeted cancer drugs have been developed to interfere with specific molecular targets, which are expected to affect the growth of cancer cells in a way that can be characterized by multi-omics data. The prediction of cancer drug sensitivity simultaneously for multiple drugs based on heterogeneous multi-omics data (e.g., mRNA expression, DNA copy number or DNA mutation) is an important but challenging task. We use joint penalized regression models for multiple cancer drugs rather than a separate model for each drug, thus being able to address the correlation structure between drugs. In addition, we employ integrative penalty factors (IPF) to allow penalizing data from different molecular data sources differently. By integrating IPF with tree-lasso, we create the IPF-tree-lasso method, which can capture the heterogeneity of multi-omics data and the correlation between drugs at the same time. Additionally, we generalize the IPF-lasso to the IPF-elastic-net, which combines ℓ_1- and ℓ_2-penalty terms and can lead to improved prediction performance. To make the computation of IPF-type methods more feasible, we present that the IPF-type methods are equivalent to the original lasso-type methods after augmenting the data matrix, and employ the Efficient Parameter Selection via Global Optimization (EPSGO) algorithm for optimizing multiple penalty factors efficiently. Simulation studies show that the new model, IPF-tree-lasso, can improve the prediction performance significantly. We demonstrate the performance of these methods on the Genomics of Drug Sensitivity in Cancer (GDSC) data.
READ FULL TEXT